Author Topic: AMD 12.1 previe driver performance (Read 23398 times)

Raistmer · « **on:** 31 Dec 2011, 09:14:56 am »

2 VLARS running:

Same 2 tasks after restart:

skildude · « **Reply #1 on:** 31 Dec 2011, 10:22:57 am »

I just check my 6970 while running 3 VHAR WU's and it looks similar.

I do have my fan speed set at 65% just to keep the temps lower than 65 C

Mike · « **Reply #2 on:** 31 Dec 2011, 10:35:53 am »

On my machine it was below 50% after restart.
Same behaviour with 11.12. Vista 64 ultimate.

Mike

Mike · « **Reply #3 on:** 31 Dec 2011, 11:10:37 am »

Here is one of Skildudes units.

http://setiathome.berkeley.edu/result.php?resultid=2239022262

7600 seconds for AR 0.29 unit is way to slow for a 6970.
The unit is marked as CPU so resheduled.

Mike

Raistmer · « **Reply #4 on:** 01 Jan 2012, 08:43:01 am »

2 non-VLAR tasks (host running w/o intervention more than day):

Same tasks after restart:

And same 2 tasks again after complete BOINC restart (no host reboot):

So, looks like BOINC responsible for such disgusting performance

Raistmer · « **Reply #5 on:** 01 Jan 2012, 08:56:45 am »

Or just weird driver behavior when it gives different results from restort to restart...

On this picture one of those 2 tasks finished (small drop). Another task started but no corresponding increase in GPU usage.
Then second task finished - great GPU usage drop, almost to zero (though 1 task still running!) and then, when both new tasks are past of initialization stage, GPU usage remains unacceptable low...

skildude · « **Reply #6 on:** 01 Jan 2012, 10:30:57 am »

Have you tried running 3 WU's at a time. See if that boosts the numbers. I'm still getting a high GPU usage from mine

Claggy · « **Reply #7 on:** 01 Jan 2012, 03:23:10 pm »

Here's an example of one of my slow running ATI_r521 AP tasks with Cat 11.12:

http://setiathome.berkeley.edu/result.php?resultid=2235201318

I've also had it with Cat 11.5 on the same host,

Strangely when the HD5770 was fitted to the E8500 and that had Cat 11.5 installed, i never saw any slowdowns, and didn't need to use the -hp switch eithier.

Claggy

Mike · « **Reply #8 on:** 01 Jan 2012, 06:02:55 pm »

I´ve noticed this also on my host with 11.5. on my driver test back in june.

Raistmer · « **Reply #9 on:** 01 Jan 2012, 06:12:30 pm »

Well, in existing apps I relied on OS/driver to cleanup GPU state when process terminated.
It's possible that new AMD drivers can't cleanup properly and need help from application.
So I will not go further in optimization and implement full OpenCL objects cleanup inside app instead. Maybe this could help with this issue.

BTW, OpenCL 1.2 is really breakthrough. It should fully support devicesplitting.
For example one can use full GPU when there are enough workitems and use only poart of GPU where too few workitems available (and we have this case in Pulse finding in VLAR for example).
Hope HD69xx will be compatible with OpenCL 1.2 too, not only HD7xxx ...

Raistmer · « **Reply #10 on:** 02 Jan 2012, 03:21:18 am »

3 VLARs running, BOINC just started

Jason G · « **Reply #11 on:** 02 Jan 2012, 04:04:44 am »

Quote from: Raistmer on 01 Jan 2012, 06:12:30 pm

Well, in existing apps I relied on OS/driver to cleanup GPU state when process terminated.
It's possible that new AMD drivers can't cleanup properly and need help from application.
So I will not go further in optimization and implement full OpenCL objects cleanup inside app instead. Maybe this could help with this issue.

Probably that'll be along the right lines. Similar (with slighlty different symptoms) was originally happening under Cuda, resolved by making sure safe exit shutdown code was used, rather than terminating the process/thread during mid kernel execution or memory transfers. The side effects on Cuda devices/drivers were more along the lines of sending the driver into failsafe/sticky-downclock, but I can see that similar mechanisms might result in resource leaks or other unpredictable problems like that.

Jason

Claggy · « **Reply #12 on:** 02 Jan 2012, 05:38:56 am »

One ATI_r521 AP task running slowly, then GPUs suspended and resummed

Claggy

skildude · « **Reply #13 on:** 02 Jan 2012, 09:28:48 am »

for giggles I decided to try 4 WU's on my 6970. All seems well

Frizz · « **Reply #14 on:** 02 Jan 2012, 09:51:50 am »

On my 6970 I get the highest RAC when running 4 Astropulse WUs in parallel. At least at times when there were such things as AP units